Pith. sign in

REVIEW 3 major objections 4 minor 50 references

ErrorEraser: Unlearning Data Bias for Improved Continual Learning

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Continual learning can be improved by deliberately forgetting memories caused by mislabeled samples, not just by preventing forgetting.

desk verdict A useful plugin for noisy labels in continual learning with a plausible mechanism, but the identification step is unvalidated and the gains may be partly regularization. read the letter →

arxiv 2506.09347 v1 pith:TSNPWCAV submitted 2025-06-11 cs.LG cs.AI

Xuemei Cao , Hanlin Gu , Xin Yang , Bingjun Wei , Haoyang Liang , Xiangkun Wang , Tianrui Li This is my paper
classification cs.LGcs.AI
keywords continuallearningmachineunlearningnoisylabelsdatabiasnormalizingflowsselectiveforgettingcatastrophicknowledgetransfer
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Continual learning models trained on streams with noisy labels absorb erroneous knowledge, and this paper argues that those errors are not a side effect but a cause of both worse retention and worse transfer to later tasks. The authors propose that a continual learner should deliberately forget such errors, and they build a plugin, ErrorEraser, that does so without being told which labels are wrong and without revisiting old-task data. ErrorEraser first fits a normalizing-flow density model to the intermediate feature space of the current task, flags the low-density tail as potentially mislabeled, then fine-tunes only those samples into a new pseudo-class and prunes that pseudo-class away. The claim is that this targeted forgetting raises accuracy and lowers forgetting across regularization-, replay-, and optimization-based CL methods, often matching or exceeding clean-data performance. A reader should care because real-world data streams contain noisy labels and errors learned early can amplify through the task sequence.

What carries the argument

The load-bearing mechanism is a normalizing-flow density model attached to the intermediate feature layer h_b of the continual-learning backbone, trained jointly with the classifier through the combined loss L = L_ce + L_ge + L_cl. The flow maps features to a multivariate normal latent distribution and gives each sample a probability density p(x); the low-density subset D_s = {x_i : p(x_i) < delta} is the flagged erroneous memory. Erasure is done by appending one pseudo-class neuron, fine-tuning the model on D_s with a pseudo-label so those samples' decision space moves into the new neuron, and then pruning that neuron and its associated activations, restoring the original output width while satisfying F_{theta'_t}(x_i) approximately 0 for the flagged samples. This machinery is what makes the claim 'forget unknown errors without historical data' operative.

What would settle it

On a benchmark with known injected noise, compute the precision and recall of the flagging rule p(x) < delta against the ground-truth noisy labels, and compare the effect of removing the flagged set versus removing the same number of random samples from clean data. If flagged sets overlap the true noisy labels no better than chance, or if pruning them hurts clean-data accuracy as much as random pruning, then the reported improvements are regularization effects rather than targeted unlearning.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that unknown noisy-label bias in old tasks can be identified and unlearned in a continual learner without any access to historical data. The identification uses the probability density of task features in a compact layer of the network, learned by a normalizing flow with a composite loss that couples density learning to the classification loss; samples whose density falls below a threshold delta are treated as representative erroneous samples. The erasure then expands the output layer with one pseudo-class neuron, fine-tunes the model so those low-density samples map to that neuron, and prunes the neuron, approximately zeroing the model's output on the erroneous samples while leaving other samples' activations essentially unchanged. Across CIFAR-10, CIFAR-100, MNIST, and WebVision with injected or real label noise, the authors report that adding ErrorEraser to EWC, MAS, LWF, DERPP, TAT, and HAT improves the harmonic-mean score S, average new-task accuracy A1, average historical accuracy A2, and lowers the forgetting rate F, with some noisy-data results surpassing clean-data baselines.

Load-bearing premise

The method assumes that mislabeled samples are exactly the low-density outliers in the learned feature space, so that pruning the pseudo-class built from those outliers erases only erroneous knowledge and does not discard rare but correct samples.

Editorial extensions

If this is right

  • Adding ErrorEraser to any of the six tested CL methods raises the combined retention-and-transfer score S and lowers the forgetting rate F under 50% label noise.
  • The plugin is agnostic to the CL objective: it works with regularization losses, replay losses, and parameter-isolation masks, so the same unlearning module can wrap a variety of continual learners.
  • Because only the low-density tail D_s is fine-tuned and then pruned, the erasure is cheap in time, memory, and model size compared with retraining or contrastive purification.
  • In several settings the noisy-data model with ErrorEraser outperforms the same CL method trained on clean data, which the authors attribute to removal of conflicting or irrelevant knowledge.
  • The incremental feature-distribution learning strategy lets the density model be updated across tasks instead of retrained from scratch, keeping downstream-task overhead low.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper never reports how many flagged samples actually match the injected noisy labels; a precision-and-recall analysis against ground-truth noise would settle whether the gains come from accurate identification or from generic regularization of low-density points.
  • If low density is a reliable corruption signal, the same density estimate could be used earlier, as a sample-weighting filter during training, potentially making the post-hoc pruning step unnecessary for some noise rates.
  • Because the method is demonstrated only in task-incremental settings with known task IDs and a multi-head classifier, a natural next test is whether a shared density model can identify mislabeled samples when task boundaries are unknown or classes overlap.
  • The pseudo-class-and-prune operation is a general selective-forgetting primitive; it may extend to removing poisoned, outdated, or privacy-sensitive instances from a trained model without full retraining.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes ErrorEraser, a plugin for continual learning (CL) that identifies and erases 'erroneous memories' caused by noisy labels. It has two modules: Error Identification, which uses normalizing flows to model the probability density of task features in a compact latent space, with a new loss L_ge that downweights low-density samples during backbone training; and Error Erasure, which selects the lowest-density samples D_s, fine-tunes them into a pseudo-class, and prunes the corresponding neuron. The authors claim that this removes erroneous knowledge and improves both retention and transfer across three types of CL methods (regularization-, replay-, and optimization-based). Experiments on synthetic noisy MNIST, CIFAR-10, CIFAR-100 and real WebVision report higher composite scores S and lower forgetting for most combinations, with comparisons to noisy-label CL methods and unlearning baselines.

Significance. If the central mechanism is correct, the paper would make a useful contribution: it reframes CL as needing not only to prevent forgetting but also to intentionally forget incorrect knowledge, and it proposes a plug-and-play method that requires no access to historical data. The empirical scope is broad (six CL backbones, three synthetic datasets, one real dataset, several noise rates and task counts), and the code is released. The resource-cost analysis (Table 5) further indicates that the plugin adds modest overhead. However, the significance is tempered by the fact that the load-bearing claim---that low-density samples are noisy and that fine-tuning them into a pseudo-class erases only erroneous knowledge---is never validated against ground-truth noisy labels. The reported gains are also consistent with a weaker mechanism in which L_ge acts as a robust-loss regularizer and the forgetting fine-tuning provides additional regularization on a high-loss subset. The paper therefore needs a direct evaluation of identification accuracy and a more careful ablation of the selection rule before its main claim can be accepted.

major comments (3)
  1. [Sect. 4.1.3, Eq. (11)] The core premise is that samples with low probability density p(x) in the NF latent space are 'potential noisy labels'. Since the synthetic noise is generated with known asymmetric mappings (Sect. 7.1), the authors can directly measure the precision/recall or overlap of the selected set D_s with the true mislabeled set. No such metric is reported. This matters because the training loss L_ge (Eq. 10) already downweights low-density features during backbone training, so the post-hoc density-based selection is partially self-referential: the model is trained to make low-density samples less influential, and those same samples are then flagged as noisy. The concrete test is to report identification accuracy on the known noise labels, and to compare the end-task performance of density-based selection with random selection and with selection based on large classification loss. Without that experiment, the paper's explanatory narrative ('accurately identify and erase erroneous memories') is unsupported, and the gains could be attributed to L_ge plus fine-tuning a high-loss subset as a robust-loss regularizer.
  2. [Sect. 5.2, Table 2 and Appendix Table 8] The abstract and Section 5.2 claim 'higher accuracy and lower forgetting rates across three types of CL methods'. The appendix data do not fully support this universal claim: on CIFAR-10 at 10% noise, DERPP+ErrorEraser has S=61.81 versus S=63.68 for the baseline DERPP, and at 30% noise it has S=57.00 versus 58.08. Similarly, HAT+ErrorEraser on CIFAR-10 shows substantially higher forgetting (F=33.23 at 50% noise) than the baseline HAT (F=1.34). The paper gives a qualitative explanation for the HAT case, but no analysis for DERPP. Moreover, the experiments are run five times and averaged, yet no standard deviations, confidence intervals, or significance tests are reported anywhere. The claim of 'significant mitigation' (Sect. 5.3, Table 4) is therefore not statistically grounded, and the universal-improvement claim is contradicted by the appendix results without discussion.
  3. [Sect. 4.2.2, Eq. (13) and Fig. 6] The erasure mechanism assumes that fine-tuning only D_s into a pseudo-class and pruning the new neuron erases erroneous knowledge while leaving correct knowledge intact. The sensitivity analysis (Fig. 6) acknowledges that increasing the forgotten proportion degrades performance because 'normal samples' are included, but it does not quantify how many correct samples are affected, nor does it report the performance of the model on the actual noisy-label subset. Since D_s is defined only by a density threshold with no validation of its label-correctness composition, the safety of the unlearning step is unquantified. A direct measurement of the false-positive rate of D_s (correct samples selected) would settle whether the procedure is truly selective or simply a high-loss fine-tuning step that improves generalization regardless of label correctness.
minor comments (4)
  1. [Throughout] There are frequent typos and formatting artifacts: 'Taks' in Tables 2, 8, and 9; concatenated numbers in Table 8 ('14.5157.90', '9.0661.12', etc.); and inconsistent spacing around equations. These should be cleaned up.
  2. [Eq. (16)] The composite metric S is introduced without a motivation or a reference to a previously established metric. The formula as printed is S = A1*A2*n / (A1 + A2*n + F_k); please confirm that the intended denominator is indeed A1 + (A2 * n) + F_k and clarify why n multiplies A2 but not A1.
  3. [Sect. 7.1] The description of WebVision says 'the top 14 largest classes' and 'seven tasks with randomly paired classes', but the main-text Table 3 shows results for WebVision without specifying the task count. Please align these details and state the noise rate and class/task split consistently.
  4. [Sect. 5.3, Table 5] In Table 5, the resource numbers for 'Cover' and the baselines are reported with order-of-magnitude values ('~4 x 10^1' seconds). It would be clearer to report exact numbers and to state which hardware and implementation were used, since the efficiency comparison is used to support the plugin's practicality.

Circularity Check

1 steps flagged · score 6.0 of 10

Error identification is self-definitional: the 'noisy label' set is constructed as the low-density set (Eq. 11), and the same density is used to downweight those samples during training, so the claim of accurate identification is a restatement of the selection rule rather than an independent prediction.

  1. self definitional [Section 4.1.3 and Eq. (11); with Eq. (10) in Section 4.1.2]
    "Considering that abnormal features relative to the current task may be caused by noisy label samples, we evaluate the probability density value of each sample's corresponding features by Eq. (2) after completing the distribution learning model. The likelihood of a sample being a noisy label is measured by the probability density value. Finally, samples with low probability density values at the distribution margins are detected, enabling the identification of potential noisy labels."

    The ground-truth noisy set D'_t is defined in Sec. 3.2 by label mismatch, but the paper never measures overlap between the selected D_s and D'_t, despite using known asymmetric noise mappings in Sec. 7.1. Instead, Eq. (11) defines D_s = {(x_i,y_i) in D_t | p(x_i) < delta}, and Sec. 4.1.3 equates 'likelihood of being a noisy label' with the density value. Thus the 'accurate identification of potentially biased samples' is not an independently tested prediction; it is the selection rule restated. Moreover, Eq. (10) already uses a density-derived weight q_Dt to downweight low-density samples during training, so the same low-density assumption is baked into the model before the later selection. The end-task test accuracy is external, so this is partial rather than total circularity.

full rationale

The paper's main empirical claim, that ErrorEraser improves accuracy and reduces forgetting on held-out test tasks, is evaluated against external test sets and is not fitted to the target, so it is not circular. No load-bearing self-citation or imported uniqueness theorem is present. However, the paper's mechanistic claim of accurate noisy-label identification is self-definitional: the identified set is, by Eq. (11), exactly the set of low-probability-density samples, and Sec. 4.1.3 states that the likelihood of a sample being noisy is measured by the density value. The true noisy labels are never used to measure precision or recall of this identification, despite the synthetic noise being generated from known label mappings. Furthermore, the same density concept is used in the L_ge loss (Eq. 10) to downweight low-density features during training, so the model is trained to discount the very samples that are later flagged as errors. This creates a partial circularity in the 'accurate identification' and 'erasing only erroneous knowledge' narrative, while leaving the observed downstream performance improvements as independent evidence. The score reflects that the central identification claim reduces by construction, but the overall empirical performance remains externally validated.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim depends on two domain assumptions (low-density equals noisy, and pseudo-class erasure is harmless) and several unreported implementation choices (delta percentile, number of forgetting epochs, flow architecture, noise placement). No new physical or architectural entities are introduced beyond the pseudo-class neuron, which is part of the method rather than an independent postulate.

free parameters (3)
  • Threshold delta for low-density selection (Eq. 11)
    Paper says delta is set based on a density percentile and includes a sensitivity study, but the exact percentile used in main experiments is not reported.
  • Number of forgetting fine-tuning epochs
    Algorithm 1 line 13 runs a fine-tuning loop; sensitivity analysis varies the number of epochs but no default value is reported.
  • Normalizing flow architecture
    Number of flow layers and hidden dimensions are not specified; the density estimate depends on this architecture, which affects which samples are selected.
assumptions (4)
  • domain assumption Mislabeled samples are outliers in the feature-space probability density estimated by the normalizing flow.
    Section 4.1.3 equates low density with noisy labels; no ground-truth validation is provided.
  • domain assumption Fine-tuning low-density samples into an added pseudo-class and pruning that neuron erases only erroneous knowledge, leaving normal samples unaffected.
    Section 4.2.2 and Eq. (13) assert F_theta'(x_i) is near zero for erased samples; no measurement on the remaining data is shown.
  • domain assumption Feature distributions of previous tasks can be represented by generated samples from the normalizing flow without revisiting old data.
    Section 4.1.1 Eq. (6)-(7) requires this for the incremental strategy; if generated features drift, identification in later tasks could be biased.
  • standard math Normalizing flow change-of-variables and Gaussian prior computations are valid.
    Eqs. (1)-(3) rely on standard normalizing flow theory; Eq. (1) as printed omits the inverse Jacobian, though Eq. (2) is the standard form.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ErrorEraser: Unlearning Data Bias for Improved Continual Learning." pith.science (2026). https://pith.science/paper/TSNPWCAV

@misc{pith2026250609347,
  author       = {Pith},
  title        = {Pith review of: ErrorEraser: Unlearning Data Bias for Improved Continual Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TSNPWCAV}},
  note         = {Machine review of arXiv:2506.09347}
}
read the original abstract

Continual Learning (CL) primarily aims to retain knowledge to prevent catastrophic forgetting and transfer knowledge to facilitate learning new tasks. Unlike traditional methods, we propose a novel perspective: CL not only needs to prevent forgetting, but also requires intentional forgetting.This arises from existing CL methods ignoring biases in real-world data, leading the model to learn spurious correlations that transfer and amplify across tasks. From feature extraction and prediction results, we find that data biases simultaneously reduce CL's ability to retain and transfer knowledge. To address this, we propose ErrorEraser, a universal plugin that removes erroneous memories caused by biases in CL, enhancing performance in both new and old tasks. ErrorEraser consists of two modules: Error Identification and Error Erasure. The former learns the probability density distribution of task data in the feature space without prior knowledge, enabling accurate identification of potentially biased samples. The latter ensures only erroneous knowledge is erased by shifting the decision space of representative outlier samples. Additionally, an incremental feature distribution learning strategy is designed to reduce the resource overhead during error identification in downstream tasks. Extensive experimental results show that ErrorEraser significantly mitigates the negative impact of data biases, achieving higher accuracy and lower forgetting rates across three types of CL methods. The code is available at https://github.com/diadai/ErrorEraser.

Figures

Figures reproduced from arXiv: 2506.09347 by the authors.

Figure 1
Figure 1. Diagram of the data bias. Although current CL methods have achieved significant progress [22, 45], they heavily rely on the quality of training data [16], of￾ten overlooking more realistic settings where data biases exist in real-world applications. During data collection, due to the large scale and the high cost of expert annotation, data often contains noisy labels, as shown in [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗
Figure 2
Figure 2. The impact of noisy label biases on feature extrac [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustrations of our ErrorEraser plugin for classifier training in CL. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: t-SNE visualization of feature extraction for Task [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The attention map of three CL methods before and [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Performance on different parameters. 6 CONCLUSION This paper explores the dual impact of noisy labels on CL and in￾troduces ErrorEraser, a universal plugin designed to address this issue. ErrorEraser learns the probability density distribution of data to effectively id…
Figure 7
Figure 7. Figure 7: Accuracy and forgetting before and after Error [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: t-SNE visualization of the extracted feature outputs [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 44 canonical work pages

  1. [1]

    Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. 2018. Memory aware synapses: Learning what (not) to forget. In Proceedings of the European conference on computer vision . 139–154

    work page 2018

  2. [2]

    Rahaf Aljundi, Klaas Kelchtermans, and Tinne Tuytelaars. 2019. Task-free con- tinual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 11254–11263. KDD ’25, August 3–7, 2025, Toronto, ON, Canada. Xuemei Cao, et al

    work page 2019

  3. [3]

    Nourhan Bayasi, Jamil Fayyad, Alceu Bissoto, Ghassan Hamarneh, and Rafeef Garbi. 2024. Biaspruner: Debiased continual learning for medical image classifi- cation. In International Conference on Medical Image Computing and Computer- Assisted Intervention. Springer, 90–101

    work page 2024

  4. [4]

    Florian Peter Busch, Roshni Kamath, Rupert Mitchell, Wolfgang Stammer, Kristian Kersting, and Martin Mundt. 2024. Where is the Truth? The Risk of Getting Confounded in a Continual World. arXiv preprint arXiv:2402.06434 (2024)

    arXiv 2024

  5. [5]

    Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. 2020. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems 33 (2020), 15920– 15930

    work page 2020

  6. [6]

    Xuemei Cao, Xin Yang, Shuyin Xia, Guoyin Wang, and Tianrui Li. 2024. Open Continual Feature Selection via Granular-Ball Knowledge Transfer. IEEE Trans- actions on Knowledge and Data Engineering 36, 12 (2024), 8967–8980

    work page 2024

  7. [7]

    Sungmin Cha, Sungjun Cho, Dasol Hwang, Honglak Lee, Taesup Moon, and Moontae Lee. 2024. Learning to unlearn: Instance-wise unlearning for pre- trained classifiers. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 38. 11186–11194

    work page 2024

  8. [8]

    Min Chen, Weizhuo Gao, Gaoyang Liu, Kai Peng, and Chen Wang. 2023. Boundary unlearning: Rapid forgetting of deep networks via shifting the decision bound- ary. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 7766–7775

    work page 2023

Show all 50 references
  1. [9]

    Tao Fan, Hanlin Gu, Xuemei Cao, et al. 2025. Ten Challenging Problems in Feder- ated Foundation Models. IEEE Transactions on Knowledge and Data Engineering (2025), 1–20

    2025

  2. [10]

    Hanlin Gu, Gongxi Zhu, Jie Zhang, Xinyuan Zhao, Yuxing Han, Lixin Fan, and Qiang Yang. 2024. Unlearning during Learning: An Efficient Federated Ma- chine Unlearning Method. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence. 4035–4043

    2024

  3. [11]

    Yiduo Guo, Bing Liu, and Dongyan Zhao. 2023. Dealing with cross-task class discrimination in online continual learning. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition . 11878–11887

    2023

  4. [12]

    Jindong Han, Hao Liu, Shui Liu, Xi Chen, Naiqiang Tan, Hua Chai, and Hui Xiong. 2023. iETA: A Robust and Scalable Incremental Learning Framework for Time-of-Arrival Estimation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 4100–4111

    2023

  5. [13]

    Alvin Heng and Harold Soh. 2023. Selective amnesia: A continual learning approach to forgetting in deep generative models.Advances in Neural Information Processing Systems 36 (2023), 17170–17194

    2023

  6. [14]

    Nazmul Karim, Umar Khalid, Ashkan Esmaeili, and Nazanin Rahnavard. 2022. Cnll: A semi-supervised approach for continual noisy label learning. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 3878–3888

    2022

  7. [15]

    Zixuan Ke, Bing Liu, Nianzu Ma, Hu Xu, and Lei Shu. 2021. Achieving forgetting prevention and knowledge transfer in continual learning. Advances in Neural Information Processing Systems 34 (2021), 22443–22456

    2021

  8. [16]

    Chris Dongjoo Kim, Jinseo Jeong, Sangwoo Moon, and Gunhee Kim. 2021. Con- tinual learning on noisy data streams via self-purified replay. In Proceedings of the IEEE/CVF international conference on computer vision . 537–547

    2021

  9. [17]

    Sein Kim, Namkyeong Lee, Donghyun Kim, Minchul Yang, and Chanyoung Park. 2023. Task Relation-aware Continual User Representation Learning. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1107–1119

    2023

  10. [18]

    Youngdong Kim, Junho Yim, Juseung Yun, and Junmo Kim. 2019. Nlnl: Negative learning for noisy labels. In Proceedings of the IEEE/CVF international conference on computer vision. 101–110

    2019

  11. [19]

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, et al . 2017. Overcoming catastrophic forgetting in neural networks. Proceedings of the na- tional academy of sciences 114, 13 (2017), 3521–3526

    2017

  12. [20]

    Jan Kremer, Fei Sha, and Christian Igel. 2018. Robust active label correction. In International conference on artificial intelligence and statistics . 308–316

    2018

  13. [21]

    Alex Krizhevsky, Geoffrey Hinton, et al. 2009. Learning multiple layers of features from tiny images. Technical report (2009)

    2009

  14. [22]

    Guannan Lai, Yujie Li, Xiangkun Wang, Junbo Zhang, Tianrui Li, and Xin Yang

  15. [23]

    Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. 1998. Gradient- based learning applied to document recognition. Proc. IEEE 86, 11 (1998), 2278– 2324

    1998

  16. [24]

    Junnan Li, Yongkang Wong, Qi Zhao, and Mohan S Kankanhalli. 2019. Learning to learn from noisy labeled data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 5051–5059

    2019

  17. [25]

    Miaomiao Li, Jiaqi Zhu, Xin Yang, Yi Yang, Qiang Gao, and Hongan Wang. 2023. Cl-wstc: Continual learning for weakly supervised text classification on the internet. In Proceedings of the ACM Web Conference 2023 . 1489–1499

    2023

  18. [26]

    Yujie Li, Xin Yang, Hao Wang, Xiangkun Wang, and Tianrui Li. 2024. Learning to Prompt Knowledge Transfer for Open-World Continual Learning. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 38. 13700–13708

    2024

  19. [27]

    Zhizhong Li and Derek Hoiem. 2018. Learning without forgetting. IEEE transac- tions on pattern analysis and machine intelligence 40, 12 (2018), 2935–2947

    2018

  20. [28]

    Sen Lin, Peizhong Ju, Yingbin Liang, and Ness Shroff. 2023. Theory on forgetting and generalization of continual learning. In International Conference on Machine Learning. 21078–21100

    2023

  21. [29]

    Sihao Lin, Hongwei Xie, Bing Wang, Kaicheng Yu, Xiaojun Chang, Xiaodan Liang, and Gang Wang. 2022. Knowledge distillation via the target-aware transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 10915–10924

    2022

  22. [30]

    Yiwei Luo and Min Jiang. 2023. Solving Continual Learning with Noisy Labels by Sample Selection and Replay. In 2023 International Joint Conference on Neural Networks. 1–8

    2023

  23. [31]

    Imad Eddine Marouf, Subhankar Roy, Enzo Tartaglione, and Stéphane Lathuilière

  24. [32]

    M Jehanzeb Mirza, Marc Masana, Horst Possegger, and Horst Bischof. 2022. An efficient domain-incremental learning approach to drive in all weather condi- tions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 3001–3011

    2022

  25. [33]

    Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. 2017. Making deep neural networks robust to label noise: A loss correction approach. In Proceedings of the IEEE conference on computer vision and pattern recognition. 1944–1952

    2017

  26. [34]

    Danilo Rezende and Shakir Mohamed. 2015. Variational inference with normaliz- ing flows. In International conference on machine learning . 1530–1538

    2015

  27. [35]

    Tim GJ Rudner, Freddie Bickford Smith, Qixuan Feng, Yee Whye Teh, and Yarin Gal. 2022. Continual learning via sequential function-space variational inference. In International Conference on Machine Learning . 18871–18887

    2022

  28. [36]

    Tomás J Ryan and Paul W Frankland. 2022. Forgetting as a form of adaptive engram cell plasticity. Nature Reviews Neuroscience 23, 3 (2022), 173–186

    2022

  29. [37]

    Joan Serra, Didac Suris, Marius Miron, and Alexandros Karatzoglou. 2018. Over- coming catastrophic forgetting with hard attention to the task. In International conference on machine learning . 4548–4557

    2018

  30. [38]

    Lianlei Shan, Wenzhang Zhou, Wei Li, and Xingyu Ding. 2024. Lifelong Learning and Selective Forgetting via Contrastive Strategy. arXiv preprint arXiv:2405.18663 (2024)

    2024 arXiv

  31. [39]

    Takashi Shibata, Go Irie, Daiki Ikami, and Yu Mitsuzumi. 2021. Learning with Selective Forgetting. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence. 989–996

    2021

  32. [40]

    Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. 2017. Continual learning with deep generative replay. Advances in neural information processing systems 30 (2017)

    2017

  33. [41]

    Hwanjun Song, Minseok Kim, Dongmin Park, Yooju Shin, and Jae-Gil Lee. 2022. Learning from noisy labels with deep neural networks: A survey.IEEE transactions on neural networks and learning systems 34, 11 (2022), 8135–8153

    2022

  34. [42]

    Zhicheng Sun, Yadong Mu, and Gang Hua. 2023. Regularizing second-order influences for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 20166–20175

    2023

  35. [43]

    Gido M Van de Ven, Tinne Tuytelaars, and Andreas S Tolias. 2022. Three types of incremental learning. Nature Machine Intelligence 4, 12 (2022), 1185–1197

    2022

  36. [44]

    Andreas Veit, Neil Alldrin, Gal Chechik, Ivan Krasin, Abhinav Gupta, and Serge Belongie. 2017. Learning from noisy large-scale datasets with minimal supervi- sion. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion. 839–847

    2017

  37. [45]

    Xin Yang, Hao Yu, Xin Gao, Hao Wang, Junbo Zhang, and Tianrui Li. 2024. Federated continual learning via knowledge fusion: A survey. IEEE Transactions on Knowledge and Data Engineering 36, 8 (2024), 3832–3850

    2024

  38. [46]

    Jaehong Yoon, Divyam Madaan, Eunho Yang, and Sung Ju Hwang. 2021. On- line coreset selection for rehearsal-based continual learning. arXiv preprint arXiv:2106.01085 (2021)

    2021 arXiv

  39. [47]

    Qing Yu and Kiyoharu Aizawa. 2020. Unknown class label cleaning for learn- ing with open-set noisy labels. In 2020 IEEE International Conference on Image Processing. IEEE, 1731–1735

    2020

  40. [48]

    Zhen Zhao, Zhizhong Zhang, Xin Tan, Jun Liu, Yanyun Qu, Yuan Xie, and Lizhuang Ma. 2023. Rethinking gradient projection continual learning: Stabil- ity/plasticity feature space decoupling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 3...

    2023

  41. [2024]

    In European Conference on Computer Vision

    Weighted ensemble models are strong continual learners. In European Conference on Computer Vision . 306–324

  42. [2025]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Order-robust class incremental learning: Graph-driven dynamic similarity grouping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.